]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blame - drivers/gpu/drm/i915/i915_gem_context.c
drm/i915: page table generalizations
[mirror_ubuntu-artful-kernel.git] / drivers / gpu / drm / i915 / i915_gem_context.c
CommitLineData
254f965c
BW
1/*
2 * Copyright © 2011-2012 Intel Corporation
3 *
4 * Permission is hereby granted, free of charge, to any person obtaining a
5 * copy of this software and associated documentation files (the "Software"),
6 * to deal in the Software without restriction, including without limitation
7 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8 * and/or sell copies of the Software, and to permit persons to whom the
9 * Software is furnished to do so, subject to the following conditions:
10 *
11 * The above copyright notice and this permission notice (including the next
12 * paragraph) shall be included in all copies or substantial portions of the
13 * Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21 * IN THE SOFTWARE.
22 *
23 * Authors:
24 * Ben Widawsky <ben@bwidawsk.net>
25 *
26 */
27
28/*
29 * This file implements HW context support. On gen5+ a HW context consists of an
30 * opaque GPU object which is referenced at times of context saves and restores.
31 * With RC6 enabled, the context is also referenced as the GPU enters and exists
32 * from RC6 (GPU has it's own internal power context, except on gen5). Though
33 * something like a context does exist for the media ring, the code only
34 * supports contexts for the render ring.
35 *
36 * In software, there is a distinction between contexts created by the user,
37 * and the default HW context. The default HW context is used by GPU clients
38 * that do not request setup of their own hardware context. The default
39 * context's state is never restored to help prevent programming errors. This
40 * would happen if a client ran and piggy-backed off another clients GPU state.
41 * The default context only exists to give the GPU some offset to load as the
42 * current to invoke a save of the context we actually care about. In fact, the
43 * code could likely be constructed, albeit in a more complicated fashion, to
44 * never use the default context, though that limits the driver's ability to
45 * swap out, and/or destroy other contexts.
46 *
47 * All other contexts are created as a request by the GPU client. These contexts
48 * store GPU state, and thus allow GPU clients to not re-emit state (and
49 * potentially query certain state) at any time. The kernel driver makes
50 * certain that the appropriate commands are inserted.
51 *
52 * The context life cycle is semi-complicated in that context BOs may live
53 * longer than the context itself because of the way the hardware, and object
54 * tracking works. Below is a very crude representation of the state machine
55 * describing the context life.
56 * refcount pincount active
57 * S0: initial state 0 0 0
58 * S1: context created 1 0 0
59 * S2: context is currently running 2 1 X
60 * S3: GPU referenced, but not current 2 0 1
61 * S4: context is current, but destroyed 1 1 0
62 * S5: like S3, but destroyed 1 0 1
63 *
64 * The most common (but not all) transitions:
65 * S0->S1: client creates a context
66 * S1->S2: client submits execbuf with context
67 * S2->S3: other clients submits execbuf with context
68 * S3->S1: context object was retired
69 * S3->S2: clients submits another execbuf
70 * S2->S4: context destroy called with current context
71 * S3->S5->S0: destroy path
72 * S4->S5->S0: destroy path on current context
73 *
74 * There are two confusing terms used above:
75 * The "current context" means the context which is currently running on the
508842a0 76 * GPU. The GPU has loaded its state already and has stored away the gtt
254f965c
BW
77 * offset of the BO. The GPU is not actively referencing the data at this
78 * offset, but it will on the next context switch. The only way to avoid this
79 * is to do a GPU reset.
80 *
81 * An "active context' is one which was previously the "current context" and is
82 * on the active list waiting for the next context switch to occur. Until this
83 * happens, the object must remain at the same gtt offset. It is therefore
84 * possible to destroy a context, but it is still active.
85 *
86 */
87
760285e7
DH
88#include <drm/drmP.h>
89#include <drm/i915_drm.h>
254f965c 90#include "i915_drv.h"
198c974d 91#include "i915_trace.h"
254f965c 92
40521054
BW
93/* This is a HW constraint. The value below is the largest known requirement
94 * I've seen in a spec to date, and that was a workaround for a non-shipping
95 * part. It should be safe to decrease this, but it's more future proof as is.
96 */
b731d33d
BW
97#define GEN6_CONTEXT_ALIGN (64<<10)
98#define GEN7_CONTEXT_ALIGN 4096
40521054 99
b731d33d
BW
100static size_t get_context_alignment(struct drm_device *dev)
101{
102 if (IS_GEN6(dev))
103 return GEN6_CONTEXT_ALIGN;
104
105 return GEN7_CONTEXT_ALIGN;
106}
107
254f965c
BW
108static int get_context_size(struct drm_device *dev)
109{
110 struct drm_i915_private *dev_priv = dev->dev_private;
111 int ret;
112 u32 reg;
113
114 switch (INTEL_INFO(dev)->gen) {
115 case 6:
116 reg = I915_READ(CXT_SIZE);
117 ret = GEN6_CXT_TOTAL_SIZE(reg) * 64;
118 break;
119 case 7:
4f91dd6f 120 reg = I915_READ(GEN7_CXT_SIZE);
2e4291e0 121 if (IS_HASWELL(dev))
a0de80a0 122 ret = HSW_CXT_TOTAL_SIZE;
2e4291e0
BW
123 else
124 ret = GEN7_CXT_TOTAL_SIZE(reg) * 64;
254f965c 125 break;
8897644a
BW
126 case 8:
127 ret = GEN8_CXT_TOTAL_SIZE;
128 break;
254f965c
BW
129 default:
130 BUG();
131 }
132
133 return ret;
134}
135
dce3271b 136void i915_gem_context_free(struct kref *ctx_ref)
40521054 137{
273497e5 138 struct intel_context *ctx = container_of(ctx_ref,
ae6c4806 139 typeof(*ctx), ref);
40521054 140
198c974d
DCS
141 trace_i915_context_free(ctx);
142
ae6c4806 143 if (i915.enable_execlists)
ede7d42b 144 intel_lr_context_free(ctx);
c7c48dfd 145
ae6c4806
DV
146 i915_ppgtt_put(ctx->ppgtt);
147
2f295791
BW
148 if (ctx->legacy_hw_ctx.rcs_state)
149 drm_gem_object_unreference(&ctx->legacy_hw_ctx.rcs_state->base);
c7c48dfd 150 list_del(&ctx->link);
40521054
BW
151 kfree(ctx);
152}
153
8c857917 154struct drm_i915_gem_object *
aa0c13da
OM
155i915_gem_alloc_context_obj(struct drm_device *dev, size_t size)
156{
157 struct drm_i915_gem_object *obj;
158 int ret;
159
160 obj = i915_gem_alloc_object(dev, size);
161 if (obj == NULL)
162 return ERR_PTR(-ENOMEM);
163
164 /*
165 * Try to make the context utilize L3 as well as LLC.
166 *
167 * On VLV we don't have L3 controls in the PTEs so we
168 * shouldn't touch the cache level, especially as that
169 * would make the object snooped which might have a
170 * negative performance impact.
171 */
172 if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev)) {
173 ret = i915_gem_object_set_cache_level(obj, I915_CACHE_L3_LLC);
174 /* Failure shouldn't ever happen this early */
175 if (WARN_ON(ret)) {
176 drm_gem_object_unreference(&obj->base);
177 return ERR_PTR(ret);
178 }
179 }
180
181 return obj;
182}
183
273497e5 184static struct intel_context *
0eea67eb 185__create_hw_context(struct drm_device *dev,
ee960be7 186 struct drm_i915_file_private *file_priv)
40521054
BW
187{
188 struct drm_i915_private *dev_priv = dev->dev_private;
273497e5 189 struct intel_context *ctx;
c8c470af 190 int ret;
40521054 191
f94982b0 192 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
146937e5
BW
193 if (ctx == NULL)
194 return ERR_PTR(-ENOMEM);
40521054 195
dce3271b 196 kref_init(&ctx->ref);
691e6415 197 list_add_tail(&ctx->link, &dev_priv->context_list);
40521054 198
691e6415 199 if (dev_priv->hw_context_size) {
aa0c13da
OM
200 struct drm_i915_gem_object *obj =
201 i915_gem_alloc_context_obj(dev, dev_priv->hw_context_size);
202 if (IS_ERR(obj)) {
203 ret = PTR_ERR(obj);
4615d4c9 204 goto err_out;
691e6415 205 }
ea0c76f8 206 ctx->legacy_hw_ctx.rcs_state = obj;
691e6415 207 }
40521054
BW
208
209 /* Default context will never have a file_priv */
691e6415
CW
210 if (file_priv != NULL) {
211 ret = idr_alloc(&file_priv->context_idr, ctx,
821d66dd 212 DEFAULT_CONTEXT_HANDLE, 0, GFP_KERNEL);
691e6415
CW
213 if (ret < 0)
214 goto err_out;
215 } else
821d66dd 216 ret = DEFAULT_CONTEXT_HANDLE;
dce3271b
MK
217
218 ctx->file_priv = file_priv;
821d66dd 219 ctx->user_handle = ret;
3ccfd19d
BW
220 /* NB: Mark all slices as needing a remap so that when the context first
221 * loads it will restore whatever remap state already exists. If there
222 * is no remap info, it will be a NOP. */
223 ctx->remap_slice = (1 << NUM_L3_SLICES(dev)) - 1;
40521054 224
676fa572
CW
225 ctx->hang_stats.ban_period_seconds = DRM_I915_CTX_BAN_PERIOD;
226
146937e5 227 return ctx;
40521054
BW
228
229err_out:
dce3271b 230 i915_gem_context_unreference(ctx);
146937e5 231 return ERR_PTR(ret);
40521054
BW
232}
233
254f965c
BW
234/**
235 * The default context needs to exist per ring that uses contexts. It stores the
236 * context state of the GPU for applications that don't utilize HW contexts, as
237 * well as an idle case.
238 */
273497e5 239static struct intel_context *
0eea67eb 240i915_gem_create_context(struct drm_device *dev,
d624d86e 241 struct drm_i915_file_private *file_priv)
254f965c 242{
42c3b603 243 const bool is_global_default_ctx = file_priv == NULL;
273497e5 244 struct intel_context *ctx;
bdf4fd7e 245 int ret = 0;
40521054 246
b731d33d 247 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
40521054 248
0eea67eb 249 ctx = __create_hw_context(dev, file_priv);
146937e5 250 if (IS_ERR(ctx))
a45d0f6a 251 return ctx;
40521054 252
ea0c76f8 253 if (is_global_default_ctx && ctx->legacy_hw_ctx.rcs_state) {
42c3b603
CW
254 /* We may need to do things with the shrinker which
255 * require us to immediately switch back to the default
256 * context. This can cause a problem as pinning the
257 * default context also requires GTT space which may not
258 * be available. To avoid this we always pin the default
259 * context.
260 */
ea0c76f8 261 ret = i915_gem_obj_ggtt_pin(ctx->legacy_hw_ctx.rcs_state,
1ec9e26d 262 get_context_alignment(dev), 0);
42c3b603
CW
263 if (ret) {
264 DRM_DEBUG_DRIVER("Couldn't pin %d\n", ret);
265 goto err_destroy;
266 }
267 }
268
d624d86e 269 if (USES_FULL_PPGTT(dev)) {
4d884705 270 struct i915_hw_ppgtt *ppgtt = i915_ppgtt_create(dev, file_priv);
bdf4fd7e
BW
271
272 if (IS_ERR_OR_NULL(ppgtt)) {
0eea67eb
BW
273 DRM_DEBUG_DRIVER("PPGTT setup failed (%ld)\n",
274 PTR_ERR(ppgtt));
bdf4fd7e 275 ret = PTR_ERR(ppgtt);
42c3b603 276 goto err_unpin;
ae6c4806
DV
277 }
278
279 ctx->ppgtt = ppgtt;
280 }
bdf4fd7e 281
198c974d
DCS
282 trace_i915_context_create(ctx);
283
a45d0f6a 284 return ctx;
9a3b5304 285
42c3b603 286err_unpin:
ea0c76f8
OM
287 if (is_global_default_ctx && ctx->legacy_hw_ctx.rcs_state)
288 i915_gem_object_ggtt_unpin(ctx->legacy_hw_ctx.rcs_state);
9a3b5304 289err_destroy:
dce3271b 290 i915_gem_context_unreference(ctx);
a45d0f6a 291 return ERR_PTR(ret);
254f965c
BW
292}
293
acce9ffa
BW
294void i915_gem_context_reset(struct drm_device *dev)
295{
296 struct drm_i915_private *dev_priv = dev->dev_private;
acce9ffa
BW
297 int i;
298
3e5b6f05
TD
299 if (i915.enable_execlists) {
300 struct intel_context *ctx;
301
302 list_for_each_entry(ctx, &dev_priv->context_list, link) {
303 intel_lr_context_reset(dev, ctx);
304 }
305
ecdb5fd8 306 return;
3e5b6f05 307 }
ecdb5fd8 308
acce9ffa 309 for (i = 0; i < I915_NUM_RINGS; i++) {
a4872ba6 310 struct intel_engine_cs *ring = &dev_priv->ring[i];
ea0c76f8 311 struct intel_context *lctx = ring->last_context;
acce9ffa 312
6689c167
MA
313 if (lctx) {
314 if (lctx->legacy_hw_ctx.rcs_state && i == RCS)
315 i915_gem_object_ggtt_unpin(lctx->legacy_hw_ctx.rcs_state);
acce9ffa 316
6689c167
MA
317 i915_gem_context_unreference(lctx);
318 ring->last_context = NULL;
acce9ffa 319 }
acce9ffa
BW
320 }
321}
322
8245be31 323int i915_gem_context_init(struct drm_device *dev)
254f965c
BW
324{
325 struct drm_i915_private *dev_priv = dev->dev_private;
273497e5 326 struct intel_context *ctx;
a45d0f6a 327 int i;
254f965c 328
2fa48d8d
BW
329 /* Init should only be called once per module load. Eventually the
330 * restriction on the context_disabled check can be loosened. */
331 if (WARN_ON(dev_priv->ring[RCS].default_context))
8245be31 332 return 0;
254f965c 333
ede7d42b
OM
334 if (i915.enable_execlists) {
335 /* NB: intentionally left blank. We will allocate our own
336 * backing objects as we need them, thank you very much */
337 dev_priv->hw_context_size = 0;
338 } else if (HAS_HW_CONTEXTS(dev)) {
691e6415
CW
339 dev_priv->hw_context_size = round_up(get_context_size(dev), 4096);
340 if (dev_priv->hw_context_size > (1<<20)) {
341 DRM_DEBUG_DRIVER("Disabling HW Contexts; invalid size %d\n",
342 dev_priv->hw_context_size);
343 dev_priv->hw_context_size = 0;
344 }
254f965c
BW
345 }
346
d624d86e 347 ctx = i915_gem_create_context(dev, NULL);
691e6415
CW
348 if (IS_ERR(ctx)) {
349 DRM_ERROR("Failed to create default global context (error %ld)\n",
350 PTR_ERR(ctx));
351 return PTR_ERR(ctx);
254f965c
BW
352 }
353
ede7d42b
OM
354 for (i = 0; i < I915_NUM_RINGS; i++) {
355 struct intel_engine_cs *ring = &dev_priv->ring[i];
356
357 /* NB: RCS will hold a ref for all rings */
358 ring->default_context = ctx;
ede7d42b 359 }
67e3d297 360
ede7d42b
OM
361 DRM_DEBUG_DRIVER("%s context support initialized\n",
362 i915.enable_execlists ? "LR" :
363 dev_priv->hw_context_size ? "HW" : "fake");
8245be31 364 return 0;
254f965c
BW
365}
366
367void i915_gem_context_fini(struct drm_device *dev)
368{
369 struct drm_i915_private *dev_priv = dev->dev_private;
273497e5 370 struct intel_context *dctx = dev_priv->ring[RCS].default_context;
67e3d297 371 int i;
254f965c 372
ea0c76f8 373 if (dctx->legacy_hw_ctx.rcs_state) {
691e6415
CW
374 /* The only known way to stop the gpu from accessing the hw context is
375 * to reset it. Do this as the very last operation to avoid confusing
376 * other code, leading to spurious errors. */
377 intel_gpu_reset(dev);
378
379 /* When default context is created and switched to, base object refcount
380 * will be 2 (+1 from object creation and +1 from do_switch()).
381 * i915_gem_context_fini() will be called after gpu_idle() has switched
382 * to default context. So we need to unreference the base object once
383 * to offset the do_switch part, so that i915_gem_context_unreference()
384 * can then free the base object correctly. */
385 WARN_ON(!dev_priv->ring[RCS].last_context);
386 if (dev_priv->ring[RCS].last_context == dctx) {
387 /* Fake switch to NULL context */
ea0c76f8
OM
388 WARN_ON(dctx->legacy_hw_ctx.rcs_state->active);
389 i915_gem_object_ggtt_unpin(dctx->legacy_hw_ctx.rcs_state);
691e6415
CW
390 i915_gem_context_unreference(dctx);
391 dev_priv->ring[RCS].last_context = NULL;
392 }
d3b448d9 393
ea0c76f8 394 i915_gem_object_ggtt_unpin(dctx->legacy_hw_ctx.rcs_state);
67e3d297
BW
395 }
396
397 for (i = 0; i < I915_NUM_RINGS; i++) {
a4872ba6 398 struct intel_engine_cs *ring = &dev_priv->ring[i];
67e3d297
BW
399
400 if (ring->last_context)
401 i915_gem_context_unreference(ring->last_context);
402
403 ring->default_context = NULL;
0009e46c 404 ring->last_context = NULL;
71b76d00
BW
405 }
406
dce3271b 407 i915_gem_context_unreference(dctx);
254f965c
BW
408}
409
2fa48d8d
BW
410int i915_gem_context_enable(struct drm_i915_private *dev_priv)
411{
a4872ba6 412 struct intel_engine_cs *ring;
2fa48d8d
BW
413 int ret, i;
414
2fa48d8d 415 BUG_ON(!dev_priv->ring[RCS].default_context);
bdf4fd7e 416
e7778be1
TD
417 if (i915.enable_execlists) {
418 for_each_ring(ring, dev_priv, i) {
419 if (ring->init_context) {
420 ret = ring->init_context(ring,
421 ring->default_context);
422 if (ret) {
423 DRM_ERROR("ring init context: %d\n",
424 ret);
425 return ret;
426 }
427 }
428 }
ecdb5fd8 429
e7778be1
TD
430 } else
431 for_each_ring(ring, dev_priv, i) {
432 ret = i915_switch_context(ring, ring->default_context);
433 if (ret)
434 return ret;
435 }
2fa48d8d
BW
436
437 return 0;
438}
439
40521054
BW
440static int context_idr_cleanup(int id, void *p, void *data)
441{
273497e5 442 struct intel_context *ctx = p;
40521054 443
dce3271b 444 i915_gem_context_unreference(ctx);
40521054 445 return 0;
254f965c
BW
446}
447
e422b888
BW
448int i915_gem_context_open(struct drm_device *dev, struct drm_file *file)
449{
450 struct drm_i915_file_private *file_priv = file->driver_priv;
f83d6518 451 struct intel_context *ctx;
e422b888
BW
452
453 idr_init(&file_priv->context_idr);
454
0eea67eb 455 mutex_lock(&dev->struct_mutex);
d624d86e 456 ctx = i915_gem_create_context(dev, file_priv);
0eea67eb
BW
457 mutex_unlock(&dev->struct_mutex);
458
f83d6518 459 if (IS_ERR(ctx)) {
0eea67eb 460 idr_destroy(&file_priv->context_idr);
f83d6518 461 return PTR_ERR(ctx);
0eea67eb
BW
462 }
463
e422b888
BW
464 return 0;
465}
466
254f965c
BW
467void i915_gem_context_close(struct drm_device *dev, struct drm_file *file)
468{
40521054 469 struct drm_i915_file_private *file_priv = file->driver_priv;
254f965c 470
73c273eb 471 idr_for_each(&file_priv->context_idr, context_idr_cleanup, NULL);
40521054 472 idr_destroy(&file_priv->context_idr);
40521054
BW
473}
474
273497e5 475struct intel_context *
40521054
BW
476i915_gem_context_get(struct drm_i915_file_private *file_priv, u32 id)
477{
273497e5 478 struct intel_context *ctx;
72ad5c45 479
273497e5 480 ctx = (struct intel_context *)idr_find(&file_priv->context_idr, id);
72ad5c45
BW
481 if (!ctx)
482 return ERR_PTR(-ENOENT);
483
484 return ctx;
254f965c 485}
e0556841
BW
486
487static inline int
a4872ba6 488mi_set_context(struct intel_engine_cs *ring,
273497e5 489 struct intel_context *new_context,
e0556841
BW
490 u32 hw_flags)
491{
e80f14b6 492 u32 flags = hw_flags | MI_MM_SPACE_GTT;
2c550183
CW
493 const int num_rings =
494 /* Use an extended w/a on ivb+ if signalling from other rings */
495 i915_semaphore_is_enabled(ring->dev) ?
496 hweight32(INTEL_INFO(ring->dev)->ring_mask) - 1 :
497 0;
498 int len, i, ret;
e0556841 499
12b0286f
BW
500 /* w/a: If Flush TLB Invalidation Mode is enabled, driver must do a TLB
501 * invalidation prior to MI_SET_CONTEXT. On GEN6 we don't set the value
502 * explicitly, so we rely on the value at ring init, stored in
503 * itlb_before_ctx_switch.
504 */
057f6a8a 505 if (IS_GEN6(ring->dev)) {
ac82ea2e 506 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, 0);
12b0286f
BW
507 if (ret)
508 return ret;
509 }
510
e80f14b6
BW
511 /* These flags are for resource streamer on HSW+ */
512 if (!IS_HASWELL(ring->dev) && INTEL_INFO(ring->dev)->gen < 8)
513 flags |= (MI_SAVE_EXT_STATE_EN | MI_RESTORE_EXT_STATE_EN);
514
2c550183
CW
515
516 len = 4;
517 if (INTEL_INFO(ring->dev)->gen >= 7)
518 len += 2 + (num_rings ? 4*num_rings + 2 : 0);
519
520 ret = intel_ring_begin(ring, len);
e0556841
BW
521 if (ret)
522 return ret;
523
b3f797ac 524 /* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */
2c550183 525 if (INTEL_INFO(ring->dev)->gen >= 7) {
e37ec39b 526 intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_DISABLE);
2c550183
CW
527 if (num_rings) {
528 struct intel_engine_cs *signaller;
529
530 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(num_rings));
531 for_each_ring(signaller, to_i915(ring->dev), i) {
532 if (signaller == ring)
533 continue;
534
535 intel_ring_emit(ring, RING_PSMI_CTL(signaller->mmio_base));
536 intel_ring_emit(ring, _MASKED_BIT_ENABLE(GEN6_PSMI_SLEEP_MSG_DISABLE));
537 }
538 }
539 }
e37ec39b 540
e0556841
BW
541 intel_ring_emit(ring, MI_NOOP);
542 intel_ring_emit(ring, MI_SET_CONTEXT);
ea0c76f8 543 intel_ring_emit(ring, i915_gem_obj_ggtt_offset(new_context->legacy_hw_ctx.rcs_state) |
e80f14b6 544 flags);
2b7e8082
VS
545 /*
546 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
547 * WaMiSetContext_Hang:snb,ivb,vlv
548 */
e0556841
BW
549 intel_ring_emit(ring, MI_NOOP);
550
2c550183
CW
551 if (INTEL_INFO(ring->dev)->gen >= 7) {
552 if (num_rings) {
553 struct intel_engine_cs *signaller;
554
555 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(num_rings));
556 for_each_ring(signaller, to_i915(ring->dev), i) {
557 if (signaller == ring)
558 continue;
559
560 intel_ring_emit(ring, RING_PSMI_CTL(signaller->mmio_base));
561 intel_ring_emit(ring, _MASKED_BIT_DISABLE(GEN6_PSMI_SLEEP_MSG_DISABLE));
562 }
563 }
e37ec39b 564 intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_ENABLE);
2c550183 565 }
e37ec39b 566
e0556841
BW
567 intel_ring_advance(ring);
568
569 return ret;
570}
571
a4872ba6 572static int do_switch(struct intel_engine_cs *ring,
273497e5 573 struct intel_context *to)
e0556841 574{
6f65e29a 575 struct drm_i915_private *dev_priv = ring->dev->dev_private;
273497e5 576 struct intel_context *from = ring->last_context;
e0556841 577 u32 hw_flags = 0;
967ab6b1 578 bool uninitialized = false;
aff43766 579 struct i915_vma *vma;
3ccfd19d 580 int ret, i;
e0556841 581
67e3d297 582 if (from != NULL && ring == &dev_priv->ring[RCS]) {
ea0c76f8
OM
583 BUG_ON(from->legacy_hw_ctx.rcs_state == NULL);
584 BUG_ON(!i915_gem_obj_is_pinned(from->legacy_hw_ctx.rcs_state));
67e3d297 585 }
e0556841 586
14d8ec54 587 if (from == to && !to->remap_slice)
9a3b5304
CW
588 return 0;
589
7e0d96bc
BW
590 /* Trying to pin first makes error handling easier. */
591 if (ring == &dev_priv->ring[RCS]) {
ea0c76f8 592 ret = i915_gem_obj_ggtt_pin(to->legacy_hw_ctx.rcs_state,
1ec9e26d 593 get_context_alignment(ring->dev), 0);
7e0d96bc
BW
594 if (ret)
595 return ret;
67e3d297
BW
596 }
597
acc240d4
DV
598 /*
599 * Pin can switch back to the default context if we end up calling into
600 * evict_everything - as a last ditch gtt defrag effort that also
601 * switches to the default context. Hence we need to reload from here.
602 */
603 from = ring->last_context;
604
ae6c4806 605 if (to->ppgtt) {
198c974d 606 trace_switch_mm(ring, to);
6689c167 607 ret = to->ppgtt->switch_mm(to->ppgtt, ring);
7e0d96bc
BW
608 if (ret)
609 goto unpin_out;
610 }
611
612 if (ring != &dev_priv->ring[RCS]) {
613 if (from)
614 i915_gem_context_unreference(from);
615 goto done;
616 }
617
acc240d4
DV
618 /*
619 * Clear this page out of any CPU caches for coherent swap-in/out. Note
d3373a24
CW
620 * that thanks to write = false in this call and us not setting any gpu
621 * write domains when putting a context object onto the active list
622 * (when switching away from it), this won't block.
acc240d4
DV
623 *
624 * XXX: We need a real interface to do this instead of trickery.
625 */
ea0c76f8 626 ret = i915_gem_object_set_to_gtt_domain(to->legacy_hw_ctx.rcs_state, false);
7e0d96bc
BW
627 if (ret)
628 goto unpin_out;
d3373a24 629
aff43766 630 vma = i915_gem_obj_to_ggtt(to->legacy_hw_ctx.rcs_state);
fe14d5f4
TU
631 if (!(vma->bound & GLOBAL_BIND)) {
632 ret = i915_vma_bind(vma,
633 to->legacy_hw_ctx.rcs_state->cache_level,
634 GLOBAL_BIND);
635 /* This shouldn't ever fail. */
636 if (WARN_ONCE(ret, "GGTT context bind failed!"))
637 goto unpin_out;
638 }
3af7b857 639
ea0c76f8 640 if (!to->legacy_hw_ctx.initialized || i915_gem_context_is_default(to))
e0556841 641 hw_flags |= MI_RESTORE_INHIBIT;
e0556841 642
e0556841 643 ret = mi_set_context(ring, to, hw_flags);
7e0d96bc
BW
644 if (ret)
645 goto unpin_out;
e0556841 646
3ccfd19d
BW
647 for (i = 0; i < MAX_L3_SLICES; i++) {
648 if (!(to->remap_slice & (1<<i)))
649 continue;
650
651 ret = i915_gem_l3_remap(ring, i);
652 /* If it failed, try again next round */
653 if (ret)
654 DRM_DEBUG_DRIVER("L3 remapping failed\n");
655 else
656 to->remap_slice &= ~(1<<i);
657 }
658
e0556841
BW
659 /* The backing object for the context is done after switching to the
660 * *next* context. Therefore we cannot retire the previous context until
661 * the next context has already started running. In fact, the below code
662 * is a bit suboptimal because the retiring can occur simply after the
663 * MI_SET_CONTEXT instead of when the next seqno has completed.
664 */
112522f6 665 if (from != NULL) {
ea0c76f8
OM
666 from->legacy_hw_ctx.rcs_state->base.read_domains = I915_GEM_DOMAIN_INSTRUCTION;
667 i915_vma_move_to_active(i915_gem_obj_to_ggtt(from->legacy_hw_ctx.rcs_state), ring);
e0556841
BW
668 /* As long as MI_SET_CONTEXT is serializing, ie. it flushes the
669 * whole damn pipeline, we don't need to explicitly mark the
670 * object dirty. The only exception is that the context must be
671 * correct in case the object gets swapped out. Ideally we'd be
672 * able to defer doing this until we know the object would be
673 * swapped, but there is no way to do that yet.
674 */
ea0c76f8 675 from->legacy_hw_ctx.rcs_state->dirty = 1;
41c52415
JH
676 BUG_ON(i915_gem_request_get_ring(
677 from->legacy_hw_ctx.rcs_state->last_read_req) != ring);
112522f6 678
c0321e2c 679 /* obj is kept alive until the next request by its active ref */
ea0c76f8 680 i915_gem_object_ggtt_unpin(from->legacy_hw_ctx.rcs_state);
112522f6 681 i915_gem_context_unreference(from);
e0556841
BW
682 }
683
ea0c76f8
OM
684 uninitialized = !to->legacy_hw_ctx.initialized && from == NULL;
685 to->legacy_hw_ctx.initialized = true;
967ab6b1 686
67e3d297 687done:
112522f6
CW
688 i915_gem_context_reference(to);
689 ring->last_context = to;
e0556841 690
967ab6b1 691 if (uninitialized) {
86d7f238 692 if (ring->init_context) {
771b9a53 693 ret = ring->init_context(ring, to);
86d7f238
AS
694 if (ret)
695 DRM_ERROR("ring init context: %d\n", ret);
696 }
46470fc9
MK
697 }
698
e0556841 699 return 0;
7e0d96bc
BW
700
701unpin_out:
702 if (ring->id == RCS)
ea0c76f8 703 i915_gem_object_ggtt_unpin(to->legacy_hw_ctx.rcs_state);
7e0d96bc 704 return ret;
e0556841
BW
705}
706
707/**
708 * i915_switch_context() - perform a GPU context switch.
709 * @ring: ring for which we'll execute the context switch
96a6f0f1 710 * @to: the context to switch to
e0556841
BW
711 *
712 * The context life cycle is simple. The context refcount is incremented and
713 * decremented by 1 and create and destroy. If the context is in use by the GPU,
ecdb5fd8 714 * it will have a refcount > 1. This allows us to destroy the context abstract
e0556841 715 * object while letting the normal object tracking destroy the backing BO.
ecdb5fd8
TD
716 *
717 * This function should not be used in execlists mode. Instead the context is
718 * switched by writing to the ELSP and requests keep a reference to their
719 * context.
e0556841 720 */
a4872ba6 721int i915_switch_context(struct intel_engine_cs *ring,
273497e5 722 struct intel_context *to)
e0556841
BW
723{
724 struct drm_i915_private *dev_priv = ring->dev->dev_private;
e0556841 725
ecdb5fd8 726 WARN_ON(i915.enable_execlists);
0eea67eb
BW
727 WARN_ON(!mutex_is_locked(&dev_priv->dev->struct_mutex));
728
ea0c76f8 729 if (to->legacy_hw_ctx.rcs_state == NULL) { /* We have the fake context */
691e6415
CW
730 if (to != ring->last_context) {
731 i915_gem_context_reference(to);
732 if (ring->last_context)
733 i915_gem_context_unreference(ring->last_context);
734 ring->last_context = to;
735 }
c482972a 736 return 0;
a95f6a00 737 }
c482972a 738
67e3d297 739 return do_switch(ring, to);
e0556841 740}
84624813 741
ec3e9963 742static bool contexts_enabled(struct drm_device *dev)
691e6415 743{
ec3e9963 744 return i915.enable_execlists || to_i915(dev)->hw_context_size;
691e6415
CW
745}
746
84624813
BW
747int i915_gem_context_create_ioctl(struct drm_device *dev, void *data,
748 struct drm_file *file)
749{
84624813
BW
750 struct drm_i915_gem_context_create *args = data;
751 struct drm_i915_file_private *file_priv = file->driver_priv;
273497e5 752 struct intel_context *ctx;
84624813
BW
753 int ret;
754
ec3e9963 755 if (!contexts_enabled(dev))
5fa8be65
DV
756 return -ENODEV;
757
84624813
BW
758 ret = i915_mutex_lock_interruptible(dev);
759 if (ret)
760 return ret;
761
d624d86e 762 ctx = i915_gem_create_context(dev, file_priv);
84624813 763 mutex_unlock(&dev->struct_mutex);
be636387
DC
764 if (IS_ERR(ctx))
765 return PTR_ERR(ctx);
84624813 766
821d66dd 767 args->ctx_id = ctx->user_handle;
84624813
BW
768 DRM_DEBUG_DRIVER("HW context %d created\n", args->ctx_id);
769
be636387 770 return 0;
84624813
BW
771}
772
773int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
774 struct drm_file *file)
775{
776 struct drm_i915_gem_context_destroy *args = data;
777 struct drm_i915_file_private *file_priv = file->driver_priv;
273497e5 778 struct intel_context *ctx;
84624813
BW
779 int ret;
780
821d66dd 781 if (args->ctx_id == DEFAULT_CONTEXT_HANDLE)
c2cf2416 782 return -ENOENT;
0eea67eb 783
84624813
BW
784 ret = i915_mutex_lock_interruptible(dev);
785 if (ret)
786 return ret;
787
788 ctx = i915_gem_context_get(file_priv, args->ctx_id);
72ad5c45 789 if (IS_ERR(ctx)) {
84624813 790 mutex_unlock(&dev->struct_mutex);
72ad5c45 791 return PTR_ERR(ctx);
84624813
BW
792 }
793
821d66dd 794 idr_remove(&ctx->file_priv->context_idr, ctx->user_handle);
dce3271b 795 i915_gem_context_unreference(ctx);
84624813
BW
796 mutex_unlock(&dev->struct_mutex);
797
798 DRM_DEBUG_DRIVER("HW context %d destroyed\n", args->ctx_id);
799 return 0;
800}
c9dc0f35
CW
801
802int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data,
803 struct drm_file *file)
804{
805 struct drm_i915_file_private *file_priv = file->driver_priv;
806 struct drm_i915_gem_context_param *args = data;
807 struct intel_context *ctx;
808 int ret;
809
810 ret = i915_mutex_lock_interruptible(dev);
811 if (ret)
812 return ret;
813
814 ctx = i915_gem_context_get(file_priv, args->ctx_id);
815 if (IS_ERR(ctx)) {
816 mutex_unlock(&dev->struct_mutex);
817 return PTR_ERR(ctx);
818 }
819
820 args->size = 0;
821 switch (args->param) {
822 case I915_CONTEXT_PARAM_BAN_PERIOD:
823 args->value = ctx->hang_stats.ban_period_seconds;
824 break;
825 default:
826 ret = -EINVAL;
827 break;
828 }
829 mutex_unlock(&dev->struct_mutex);
830
831 return ret;
832}
833
834int i915_gem_context_setparam_ioctl(struct drm_device *dev, void *data,
835 struct drm_file *file)
836{
837 struct drm_i915_file_private *file_priv = file->driver_priv;
838 struct drm_i915_gem_context_param *args = data;
839 struct intel_context *ctx;
840 int ret;
841
842 ret = i915_mutex_lock_interruptible(dev);
843 if (ret)
844 return ret;
845
846 ctx = i915_gem_context_get(file_priv, args->ctx_id);
847 if (IS_ERR(ctx)) {
848 mutex_unlock(&dev->struct_mutex);
849 return PTR_ERR(ctx);
850 }
851
852 switch (args->param) {
853 case I915_CONTEXT_PARAM_BAN_PERIOD:
854 if (args->size)
855 ret = -EINVAL;
856 else if (args->value < ctx->hang_stats.ban_period_seconds &&
857 !capable(CAP_SYS_ADMIN))
858 ret = -EPERM;
859 else
860 ctx->hang_stats.ban_period_seconds = args->value;
861 break;
862 default:
863 ret = -EINVAL;
864 break;
865 }
866 mutex_unlock(&dev->struct_mutex);
867
868 return ret;
869}